TerraKeepKeeping Earth within its limits

Trust — what the AI may publish on its own, and how to challenge it

Earth Watch findings are published by AI, automatically, and checked afterwards. This page says what that means in practice. The rules are code, not promises: they live in [src/ledger/schema.ts](src/ledger/schema.ts) (checkAppend, publishGates), and the ledger refuses any entry that breaks them. Anyone can re-check the whole history with earthdeck ledger verify.

Current policy: 2026-09-26-autonomous.

What the AI may publish on its own

A finding goes public only if all of these hold:

  1. Evidence. It was opened with evidence a third party can fetch again (scene IDs,

dates, method and version).

  1. A second, independent signal. A different sensor, a different provider, or a later

pass agreed. This step is decided by fixed rules, never by an AI's opinion.

  1. A written account. An AI wrote a plain-language narration, and every claim in it

points at evidence the finding actually holds.

  1. A second pair of eyes. After that narration, a *different* reviewer — another AI

model or a person — read it and gave the verdict publish. The AI that wrote it cannot approve it. If the text is rewritten, it needs a fresh review. The latest review saying *hold* or *reject* stops publication.

  1. Tier 0–2 only. The most serious findings (tier 3) are published only by a named

person, never by the AI alone.

  1. A record of the gates. The publish entry states who wrote the narration, who

approved it, how long the party had been notified, and which policy applied. The ledger checks those claims against its own history; a publish that claims a review it doesn't have is refused.

Reviewer identities are the names written into the ledger (model:<id>, reviewer:<handle>). They are self-declared, so the separation between writer and reviewer is a recorded, checkable practice — not a cryptographic proof. Every published finding says it was AI-drafted and names the models involved.

Naming a company or institution

Findings are about assets, places and institutions — never private individuals. There is no way to record a person as a subject or party; the schema has no such field.

To name a responsible party (an operator, an owner):

Energy Monitor ownership data), and the finding must be tier 2 or higher;

AI that wrote the narration) must sign off on the attribution;

or the notice records that the party could not be reached — and "could not be reached" is itself a record, not a claim: it must list at least two channels that were actually tried, with timestamps;

How to challenge a finding

recorded word for word and shown next to the finding. "No response" is also shown.

is deleted: the finding, the retraction and the reason stay in the public ledger.

counted, so anyone can see how often we are wrong.

a retraction, but they never publish or suppress a finding on their own. The evidence and the rules above decide.

Changing these rules

Each publish entry names the policy it was made under. Changing the rules means a new policy version, so every past decision can still be read against the rules that applied at the time. A named person can still publish under the original human-review rule (one human approval at tier 1, two distinct ones at tier 2 and above, no recorded gates) — that is how findings from before this policy were published, and they remain valid. The party-notice and tier-3 rules above apply on both routes.